On the Illumination Influence for Object Learning on Robot Companions

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Collaborative robots require effective perception of everyday objects.
• Environmental factors, like illumination changes and sensor variability, significantly impact robotic visual perception and object recognition.
• Deep Convolutional Neural Networks (CNNs) are susceptible to illumination variations in object recognition tasks.

Purpose of the Study:

• To present data augmentation techniques for object recognition that enhance deep learning architectures.
• To improve deep learning-based object recognition by incorporating linear and non-linear illumination models and feature concatenation.
• To enable more realistic Human-Robot Interaction (HRI) scenarios using minimal training data and incremental interactive object learning for long-term, location-independent learning in unshaped environments.

Main Methods:

• Model-based analysis to understand the impact of illumination changes on CNN-based object recognition.
• Development and application of data augmentation techniques, including simple linear and non-linear illumination models and feature concatenation.
• Evaluation of the proposed methods' effectiveness across various training set sizes.

Main Results:

• Changes in illumination were shown to affect CNN-based object recognition approaches.
• Data augmentation successfully modified the system for more robust recognition without the need for network retraining.
• Simple brightness change models improved recognition performance across all tested training set sizes.

Conclusions:

• Data augmentation, particularly using illumination models and feature concatenation, significantly enhances the robustness of deep learning-based object recognition in robotics.
• The proposed methods facilitate more efficient and effective object learning for robots, crucial for realistic HRI in diverse environments.
• This approach allows for improved robotic perception even with limited training data, paving the way for more adaptable and intelligent robotic systems.